Embed Size (px)
Citation preview
1
An Empirical Study of Housing Wealth and Consumption
in China
by Xiaogao Li
(8673946)
Major Paper presented to the
Department of Economics of the University of Ottawa
in partial fulfillment of the requirements of the M.A. Degree
Supervisor: Professor Yongjing Zhang
ECO 6999
Ottawa, Ontario
June 2019
2
Abstract
Using the most recent data from 2006 to 2018, I examine the relationship between housing
wealth and residents’ consumption with macro models and micro models. Splitting the sample
into two periods, the macro data analysis suggests that the housing wealth effect is
significantly positive in the period from 2006 to 2012 while it is not significant from 2013 to
2018. The micro evidence shows that the housing wealth effect on household consumption is
generally positive, differing from regions, years, gender, education and other factors. The
micro analysis results can explain part of the reason why housing wealth effect is weakening
in China, but more evidence is needed to get a better understanding of this issue.
3
Table of Contents
1.Introduction.................................................................................................................................................. 1
2. Literature Review......................................................................................................................................... 4
3. Methods .................................................................................................................................................... 13
3.1 The macro model............................................................................................................................. 13
3.2 The micro model.............................................................................................................................. 17
4. Macro evidence of housing wealth effect ................................................................................................. 18
4.1 Sample data ..................................................................................................................................... 18
4.2 General tests and estimation .......................................................................................................... 20
4.2.1 Unit root tests ....................................................................................................................... 20
4.2.2 Cointegration tests ............................................................................................................... 22
4.2.3 Collinearity autocorrelation and heteroskedasticity ............................................................ 22
4.2.4 Estimation results and analysis ............................................................................................ 24
4.3 Subsample analysis (2006Q1–2012Q4) ........................................................................................... 28
4.4 Subsample analysis (2013Q1–2018Q2) ........................................................................................... 33
5. Micro evidence of housing wealth effect .................................................................................................. 36
5.1 Sample data and tests ..................................................................................................................... 36
5.2 Estimation of equations and analysis .............................................................................................. 37
5.3 Robustness test: house owners and the houseless ......................................................................... 40
6. Conclusion ................................................................................................................................................. 41
Reference ...................................................................................................................................................... 43
Appendix ....................................................................................................................................................... 46
1
1.Introduction
For a long time, investment and exports have been the main drivers of China's economic
growth, and the contribution rate of consumption was relatively low. Today, China is facing a
new economic environment, both at home and abroad. Domestically, after 30 years of
extensive economic development characterized by high input, inefficiency, industry
overcapacity, serious waste of resources and heavy environmental pollution, the cost of
production factors continues to rise, leading to a limited effect of investment on the role of
economic growth in the future. Abroad, trade protection is becoming quite prevalent. The
emergence of trade barriers increases the uncertainty of China's exports. In the future, China's
exports are likely to encounter more and more trade barriers, which can be a problem due to
its export size and trade surpluses. The dual challenges of internal and external aspects make
China's traditional economic growth model unsustainable. Therefore, expanding consumption
seems to be the best option to achieve sustainable economic growth.
Table1.1: The structure and trend of household assets (2013, 2015, 2017)
Proportion of total assets
Year 2013 2015 2017
Housing assets 68.3 70.1 77.7
Financial assets 10.3 14.6 11.8
Industrial and
commercial assets
9.9 6.9 5.6
Other assets 11.5 8.3 4.9
Note: data from CHFS（China Household Finance Survey）
In recent years, the sustained and rapid growth of house prices in China has made real estate
2
account for a high proportion of total household assets. CHFS (China Household Finance
Survey) data show that in 2013-2017, housing assets accounted for a rising share of total
assets, reaching 77.7% in 2017; financial assets accounted for less than 15%; and the share of
industrial and commercial assets and other types of assets showed a downward trend. This
shows that in China, housing assets accounts for the highest proportion of total household
assets while the proportion of financial assets is low, and the gap is increasing, which
indicates that housing assets are the main component of China's household wealth.
Figure1.1: Household financial assets allocation ratio in different countries
Data source: China urban household wealth health report
Compared with developed countries such as the United States, China's household financial
assets allocation ratio is very low, and the proportion of housing assets is very high. In 2017,
the proportion of housing assets in China's total household assets was 77.7%, much higher
than 34.6% in the United States; while the share of financial assets in China was only 11.8%,
compared with 42.6% in the United States. Meanwhile, Japan's financial assets accounted for
61.1% of total assets; Singapore, Switzerland and the United Kingdom accounted for a
61.1% 56.0% 54.4%
52.2% 48.6%
42.6% 39.8%
11.8%
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
Japan Singapore Switzerland UK Canada USA France China
3
relatively low share of financial asset allocation, but were also more than half; Canada was
48.6%; and France was lower, at 39.8%, but all of them were well above 11.8% in China. A
higher proportion of housing assets absorbs too much household liquidity and squeezes the
allocation of financial assets to households.
The termination of the distribution system of housing in China in 1998 opened the era of
housing commercialization. Then China's real estate marketization reform launched in a
comprehensive manner. After a rapid development of the real estate industry in two decades,
it has gradually become one of the pillar industries of the national economy. Along with the
prosperity of the real estate market, house prices rose rapidly from 2053 Yuan in 1999 to 7827
yuan in 2017. The rise in house prices over the past more than 20 years is a double-edged
sword. On the one hand, it brought wealth appreciation to house owners; on the other hand, it
also consumed a big part of residents ' savings and dragged the economy.
The real estate industry has a close relationship with the national economy and the stability of
the real estate market is very important for the stable development of the economy. The 2008
subprime crisis in the United States is a famous example where the whole story started from
the real estate market, and then the financial crisis caused by the real estate industry
fluctuations swept the world, resulting in financial shocks in many countries. Now in China,
the study of real estate wealth effect has some new meanings. On the one hand, the
consumption rate of residents is still low; on the other hand, China is changing into a new
4
development pattern of the economy with consumption structure upgrading. Studying the
effect of real estate wealth in the current period is of great significance for boosting
consumption and stabilizing the economy in China.
The impact of real estate price fluctuation on consumption, that is, the real estate wealth effect,
has been found to be significant in many countries around the world. So, what about the
situation in China? Dose the rise of house price crowd out consumption or boost consumption?
And how should we measure the extent of the impact? The contributions of this article can be
concluded as the following:(1) using the latest macro data and micro data to examine the
housing wealth effect in China, I get the result that the housing wealth effect is weakening in
China; (2)I analysis the reason behind this phenomenon and get some possible explanations.
2. Literature Review
The early empirical studies of housing wealth effect mainly focus on four aspects: first, test of
the existence of housing wealth effect and the measurement of the effect strength; second, the
comparison of the wealth effect between real estate and other assets, mainly financial assets;
third, the differences of housing wealth effect of different countries; and other further studies.
Early studies of the test and measurement of housing wealth effect is based on the overall
market and the entire population, but the conclusions are very controversial. An early study
5
made by Elliot (1980), which was based on a permanent income model of consumer
expenditures, chose consumers’ measurable income and asset stocks as explanatory variables
and found that cash and financial assets could affect consumer spending, while savings and
real assets have no impact on consumption. Using the behavioral life-cycle savings model,
Levin (1998) found that consumption spending is not very sensitive to changes in the value of
houses, which means the housing wealth effect does not exist.
However, some other scholars hold an opposite view. Peek (1983) questioned previous
research data, arguing that real estate accounts for a large proportion of total household assets,
so that the wealth effect of net capital gains should not be overlooked. The results of his study
showed that there was a significant housing wealth effect. Cheney (2006) made an empirical
study based on Sweden's quarterly data from 1980 to 2004, confirming that rising house
prices can increase consumer spending, with a long-term housing wealth effect of 0.11.
Some scholars have further suggested that since real estate can be both necessity goods and
investment goods in the meantime, rising house prices will have both positive and negative
effects on consumption (Cheng, 2008). Some scholars carried out researches to study the
relationship between financial environment and housing wealth effect, generally finding that
the more mature the financial industry is, the more significant the housing wealth effect would
be. Muellbauer (2007) found that the wealth effect of the housing market was more
pronounced in OECD countries with looser credit conditions. Cheng (2008) did an empirical
6
test of VAR model and found that the wealth effect of housing market will increase with the
increase of market leverage.
For the part of the comparison of the wealth effect between real estate and others, factors
including the liquidity of assets, the expected rate of return, the financial policies, asset
structure, consumers’ consumption habits, and preferences are believed to have impacts on the
wealth effect of assets.
Some scholars believe that the wealth effect of real estate is small. For example, Dvornak and
Kohler (2003) found that there was both significant equity wealth effect and housing wealth
effect in Australia, but the wealth effect of equities was stronger in the long run, where the
consumption growth generated by 1 dollar in stock wealth increased by 3 to 6 cents more than
that of real estate.
However, more studies have found that the housing wealth effect is more pronounced. Some
scholars use the time series data of one country to make empirical analysis and find that under
the same conditions, the housing wealth effect is much greater than the wealth effect of
financial assets. For example, Carroll (2018) used quarterly data of the United States to
establish a cointegration model and found that the wealth effect of real estate is about 0.09,
and the wealth effect of financial assets is only about 0.04.
7
There are also many scholars who use data from multiple countries to build panel models, and
the conclusions also support this view. For example, Bayoumi and Edison (2003) used data
from 16 countries in 30 years and Case (2005) built a panel of 14 countries observed annually
during 25 years to test the relationship between housing wealth, financial wealth, and
consumer spending. The results show that countries generally have significant and positive
housing wealth effect, which is larger than the stock market wealth effect.
Because of many factors that influence asset wealth effect, the comparison of housing wealth
effect in different countries has also aroused lots of attention. Alexander (2002) makes an
empirical analysis using 1985-2000 panel data from 16 OECD countries and find that the
housing wealth effect is 0.035, where the number of market-led countries is 0.031 and that of
the bank-led countries is 0.107. Similarly, Catte (2004) uses OECD countries’ panel data as
sample and find that the housing wealth effect varies greatly from country to country: France
and Germany are not significant; Italy, Japan, Spain, the United Kingdom, the US and other
five countries are between 0.01 and 0.02; Australia, Canada, The Netherlands and other three
countries are between 0.05 and 0.08. Sierminska (2007) uses micro data for empirical testing
and the results show that the elasticity of house prices to consumption in Canada is about 0.12,
0.10 in France, and 0.13 in Finland. Ludwig and Sløk (2002), using data from 16 developed
countries for 1970-2000, conclude that changes in wealth in market-based economies have
resulted in greater wealth effects than bank-based economies; the wealth effect of rising house
8
prices is greater than the wealth effect of rising stock prices; over time, the wealth effects of
rising stocks and housing prices are increasing in all economies.
At present, more measurement methods have been introduced into the housing wealth effect
research and the research direction is more detailed. Diaz and Luengo-prado (2010) use a
general equilibrium model of heterogeneous agents with idiosyncratic uncertainty and
introduce the factors such as housing liquidity and collateral credit to assess the impact of the
liquidity of housing assets on wealth inequality. Kiyotaki (2011) develops a life-cycle model
of a production economy to make a quantitative theoretical research on the interaction
between house prices, aggregate production and household behavior in the life cycle of
residents. In particular, it is worth mentioning Carrol (2011), who focuses on household
consumption and savings decisions, introduces a new method based on the consumption habit
factor or the stickiness of consumption for measuring wealth effects to distinguish between
immediate and eventual wealth effects.
After the housing reform and the first round of rapid rise in house prices, Chinese scholars
began to pay attention to the study of housing wealth effect. Due to the lack of data, the early
research is mainly theoretical and qualitative, such as Tang Jianwei (2004) and Liu Jianjiang
(2005), which are a comparative study of housing and stock wealth effects and a theoretical
research on the transmission mechanism of housing wealth effects. At present, researches in
China mainly focus on the following two aspects:
9
First, for the existence of housing wealth effect in China, there is a lot of controversy about
whether there is a housing wealth effect. Some studies believe that rising house prices in
China will inhibit consumption and the wealth effect is weak or negative. For example, Liu
(2007) argues that the housing market supply structure is unreasonable and a large number of
housing demand has not been met, so the rise in house prices will only lead to the result that
residents spend more money to buy a house, instead of an increase in consumption; Luo
Zuoyan (2007) believes that the strong precautionary saving motivation and liquidity
constraints of Chinese residents have a strong limit on the effect of housing wealth. Some
scholars use provincial panel data further verified that the housing wealth effect is negative
(Jin Tao, Jiang Kai, 2012). On the other hand, the opposite view is that China's housing
wealth effect is remarkable. For example, Wei (2007) believes that China's long-term and
short-term housing wealth effect is significantly positive; Li and Shen (2007) make an
empirical study on the relationship of housing prices and consumption out of the data from
four major cities in China, and the conclusion is significantly positive.
The comparative study of the effects of stock and housing wealth in China began with the
simultaneous growth in share prices and house prices in 2007. Since then, quite a number of
studies have been made in this area while there is still no widely accepted conclusion. Some
scholars believe that China's stock market is not stable, so the housing wealth effect is more
significant than the wealth effect of stocks (Zhao Xiaoli 2007). Some scholars believe that the
10
excessive rise in house prices and liquidity constraints will inhibit the consumption demand of
residents, so the housing wealth effect is slightly weaker than wealth effect of stock (Luo,
2007). Ding Pan and Hu Zongyi (2008) believe that neither is significant, but the impact of
house price fluctuations on consumption is slightly greater.
The change of house price could affect household spending decisions by changing the savings
plan for potential buyers, or by changing a household's future income as well as spending
expectations. Many researchers have described how house price changes affect consumption
in terms of liquidity and credit constraints. In addition to affecting residents’ budgetary
constraints and borrowing constraints, changes in asset prices also affect consumption through
factors such as future expectations and the degree of development of financial markets.
The six different mechanisms of housing wealth effect proposed by Ludwig and Slok (2002)
are considered to be the most comprehensive:
1. Realized wealth effect, which means that the revenue got from the rental and sale of houses
increases by the rise in house prices, thus increasing consumption.
2. Unrealized wealth effect. Even if it does not realize, rising house prices make residents who
own houses feel richer, thus increasing consumption.
3. Liquidity constraints effect, which is mainly about housing mortgage income.
4. Budget constraint effect. For renters, rising house prices are likely to trigger the rise of
11
rents. As a result, they may reduce spending in other areas.
5. Substitution effect. For planned buyers, rising house price require more investment in
buying a house, and in the meantime they may reduce consumption.
6. Confidence effect, which is mainly reflected in the confidence of residents towards the
macro economy and the housing market. Some homeowners believe that their real estate
investment income is lasting and stable, consumption increased accordingly.
Liu Yufei (2018) proposed a two-stage transmission mechanism of housing wealth effect in
China, with 8 mechanisms in total. After the rise of house prices:
1. Residents who own housing will increase consumption due to the direct wealth effect.
2. Residents who own housing have an optimistic expectation of asset appreciation and the
macroeconomic situation, which will lead to an increase in consumer consumption. On the
other hand, the optimistic expectation of the economy may also make the residents without
houses spend more on consumption.
3. The value added of housing can alleviate the liquidity constraints of homeowners and
increase consumption; while for renters, the liquidity constraint effect is enhanced and
consumption is reduced.
4. Higher house prices lead to higher rents, which will reduce consumption of renters.
5. As a result of the custom formation effect, residents in China rarely mortgage their homes
to increase consumption, so rising house prices will not only not ease the liquidity constraints
12
of residents.
6. As a result of the intergenerational effect of consumption, rising house prices have deterred
older residents from enjoying the benefits of asset appreciation. Instead they have to consider
buying houses for their children, then the incentive for preventive savings increase.
7. The rise in house prices is often accompanied by the government's housing price
overheating regulation policy. Due to the impact of policy information impact, the
consumption of residents may be affected but the degree of impact is highly uncertain.
8. The government’s policy on house prices makes it impossible for residents to make a clear
judgment on the future trend of house prices. Thus the uncertainty is increased and the
motivation of preventive savings is increased, and consumption may decrease.
In housing markets with different sizes, different degrees of development, different degrees of
real estate liquidity and different financial market development degrees, these effects play in a
different way, so the overall wealth effect of the rise in housing prices is not clear. The time of
China's housing market development is still short with features like a gradually deepening
understanding of residential wealth and large regional difference. Consequently, the housing
wealth effect still needs more empirical test, especially for developing markets like China.
In order to capture the essentials of theories and examine the relationship between
consumption and housing wealth in more recent times, from 2006-2018, the following
hypotheses should be made:
13
Hypothesis 1: There is no significant housing wealth effect in China, which means that there
is no relationship between consumption and housing wealth.
Hypothesis 2: The change of housing wealth has significant impact on consumption and an
appreciation of housing assets will increase residents’ consumption.
Hypothesis 3: The change of housing wealth has significant impact on consumption and an
appreciation of housing assets will reduce residents’ consumption.
3. Methods
3.1 The macro model
The Life Cycle Hypothesis (LCH) of consumption and savings was set up by Franco
Modigliani, R. Brumberg and Alberto Ando (1954, 1963). According to LCH, it is believed
that rational consumers arrange their consumption and savings based on their lifetime income
to make income and consumption equal in lifetime. The family income includes labor income
and property income, so the consumption function for a household is:
lnC = a ∙ lnWR + 𝑐 ∙ ln𝑌𝐿 (1)
0<a<1, 0<c<1
where C is consumption, WR is property income, YL is labor income, “a” is the marginal
propensity to consume (MPC) of property income, “c” is the MPC of labor income. Based on
the LCH above, Modigliani established the following aggregate consumption function:
ln𝐶𝑡 = 𝑏1ln𝑌𝑡 + 𝑏2ln𝑌∗ + 𝑏3ln𝐴𝑡 (2)
14
In equation (2), 𝐶𝑡 , 𝑌𝑡 , 𝑌∗ , 𝐴𝑡 represent current consumption, current income, future
income and current property respectively; 𝑏1, 𝑏2, 𝑏3 represent the MPC of current income,
future income and current property, respectively. Modigliani believed that future income
could be counted as a multiple of current income, that is,
ln𝑌∗ = βln𝑌𝑡 (3)
The aggregate consumption function can be changed to
ln𝐶𝑡 = (𝑏1 + 𝑏2𝛽)ln𝑌𝑡 + 𝑏3ln𝐴𝑡 (4)
So at time t, consumption ln𝐶𝑡 is a function of income ln𝑌𝑡 and assets ln𝐴𝑡. LCH holds
that consumers will choose the current consumption level rationally based on their lifetime
wealth, to optimize and smooth their consumption of a lifetime. Real estate, as a form of
wealth, has naturally become an important factor affecting the consumption level of residents
along with cash, bank deposits, securities and other assets, as well as expected future income.
Considering the evidence found by CHFS (as mentioned in table 1.1) that the housing wealth
accounts for nearly 80% of households’ wealth in China, and that the proportion is in an
ascending trend, it is probably safe to substitute assets ln𝐴𝑡 to housing assets ln 𝐻𝑡, so the
estimating function form can be expressed in logarithmic form as:
ln 𝐶𝑡 = (𝑏1 + 𝑏2𝛽) ln 𝑌𝑡 + 𝑏3 ln 𝐻𝑡 + ∑ 𝑏𝑖+3 ln 𝑋𝑖,𝑡𝑛𝑖=1 + 𝜔𝑡 (5)
where 𝐶𝑡 is the aggregate amount of household consumption at time t, 𝐻𝑡 is the aggregate
value of real estate assets in RMB, 𝑋𝑖,𝑡 represents different factors that affect household’s
consumption, including savings, debts, years of education and so on. Since the value of real
15
estate assets 𝐻𝑡 equals to the average real estate price 𝐻𝑃𝑡 multiplied by the real estate
stock area 𝐻𝑆𝑡, that is:
𝐻𝑡 = 𝐻𝑃𝑡 ∗ 𝐻𝑆𝑡 (6)
So：
ln 𝐶𝑡 = (𝑏1 + 𝑏2𝛽) ln 𝑌𝑡 + 𝑏3 ln 𝐻𝑃𝑡 + 𝑏3 ln 𝐻𝑆𝑡 + ∑ 𝑏𝑖+3 ln 𝑋𝑖,𝑡𝑛𝑖=1 + 𝜔𝑡 (7)
Lag 1:
ln 𝐶𝑡−1 = (𝑏1 + 𝑏2𝛽) ln 𝑌𝑡−1 + 𝑏3 ln 𝐻𝑃𝑡−1 + 𝑏3 ln 𝐻𝑆𝑡−1 + ∑ 𝑏𝑖+3 ln 𝑋𝑖,𝑡−1𝑛𝑖=1 + 𝜔𝑡−1 (8)
Then equation (7) minus (8), we can get：
𝑐𝑡 = (𝑏1 + 𝑏2𝛽)𝑦𝑡 + 𝑏3ℎ𝑝𝑡 + 𝑏3ℎ𝑠𝑡 + ∑ 𝛼𝑖𝑥𝑖,𝑡𝑛𝑖=1 + 𝜔𝑡 − 𝜔𝑡−1 (9)
where,
𝑐𝑡 = ln 𝐶𝑡 − ln 𝐶𝑡−1 ≈ (𝐶𝑡 − 𝐶𝑡−1)/𝐶𝑡 (10)
where 𝑐𝑡 represents the rate of change of aggregate consumption, similar for other variables.
According to 2018 China Statistical Yearbook, the aggregate stock housing area is 42.02
billion square meters in China, and the quarterly housing sales area are about 0.3 billion
square meters in 2018, so the growth rate is around 0.7% per quarter, which is a relatively
small number compared with 2.2% of quarterly house price growth rate and 8% to 10%
annually income growth rate. Considering that the rate of change of other variables is
relatively small comparing to that of house prices and income, so assume that:
ℎ𝑠𝑡 ≈ 0
16
𝑥𝑖,𝑡 ≈ 0
Therefore, the approximate estimation form of equation (9) is
𝑐𝑡 = (𝑏1 + 𝑏2𝛽)𝑦𝑡 + 𝑏3ℎ𝑝𝑡 + 𝜀𝑡
Then do some substitution to make the coefficients more concise,
𝑐𝑡 = 𝛼0ℎ𝑝𝑡 + 𝛼1𝑦𝑡 + 𝜀𝑡 (11)
where 𝑦𝑡 is the rate of change of income, 𝜀𝑡 = 𝜔𝑡 − 𝜔𝑡−1. By estimating 𝛼0, we can draw
the short-term elasticity of consumption spending to changes in house price, and thus estimate
the short-term housing wealth effect.
To make a regression of time series variables, first we need to examine the stationarity
properties of the time series. When a variable has unit root, it is considered to be
non-stationary, which could lead to spurious results in time-series regression. Phillips and
Perron’s test and Augmented Dickey-Fuller test are usually used in unit root test. If the time
series are stationary, then we can consider further tests like heteroskedasticity and
autocorrelation test and start to establish suitable regression model; if the time series are
non-stationary, we can consider cointegration regression. Cointegration analysis makes it
possible to identify long-run economic relationships between two or more non-stationary
variables and to avoid the risk of spurious regression. Set y𝑡 = (y1𝑡, y2𝑡, … , y𝑘𝑡)′ as
k-dimensional random time series, t=1, 2, …,T, y𝑡~I(1), so consider the VAR(p) model as:
y𝑡 = 𝐴1y𝑡−1 + 𝐴2y𝑡−2 + ⋯ + 𝐴𝑝y𝑡−𝑝 + 𝜇𝑡 (12)
Within the Johansen multivariate cointegration framework, the following Vector Error
17
Correcting Mechanism (VECM) system is estimated after cointegration transformation:
∆y𝑡 = Γy𝑡−1 + ∑ 𝜃𝑗∆y𝑡−𝑗𝑝−1𝑗=1 + 𝜇𝑡 (13)
where, Γ = ∑ 𝐴𝑗𝑝𝑗=1 − 𝐼, 𝜃𝑗 = − ∑ 𝐴𝑗
𝑝𝑗=1 . Set Γ = αβ′, where Γ are (k×k) and α, β are
(r×r), rank(Γ)= rankα)= rank(β)=r, and β′y𝑡−1 ∼ 𝐼(0). Make a substitution in equation (13)
and get:
∆y𝑡 = αβ′y𝑡−1 + ∑ 𝜃𝑗∆y𝑡−𝑗𝑝−1𝑗=1 + 𝜇𝑡 (14)
where β′y𝑡−1 is the Error Correcting Term, which reflects the long run relationship between
variables; β contains the r cointegration relationships; α carries the corresponding adjustment
coefficients in each of the r vectors.
Another method on constructing VECM is Engle-Granger two-step method: first, carry out
cointegration regression using MLE or FMOLS method, test the cointegration relationship
between variables, and estimate the cointegration vector (long-term equilibrium relation
parameters); then, if the cointegration exists, add the residual difference obtained by the first
step to the error correction model and estimate parameters with OLS method. In this way, we
can obtain the long-term elasticity of consumption spending to changes in house price to
estimate the long-term housing wealth effect.
3.2 The micro model
According to equation (5), we can get the direct estimation model as
ln𝐶𝑖 = 𝛼0 + 𝛼1 ln 𝐻𝑖 + 𝛼2 ln 𝑌𝑖 + ∑ 𝛼𝑘,𝑖𝑋𝑘,𝑖𝑚𝑘=1 + 𝜀𝑖 (15)
ln𝐶𝑖 , ln 𝐻𝑖 and ln 𝑌𝑖 represent the logarithm of consumption, house value and income of
18
household “i” at the time when the survey is conducted. X represents control variables
including age, gender, level of education, family population, source of housing, survey year,
regional factors and so on. By estimating 𝛼1, we can get the elasticity of consumption to
house value as an estimate of micro housing wealth effect. In order to further test the housing
wealth effect with differences in genders, ages, years and regions, I construct the combined
term of virtual variables and house value to estimate time characteristics, regional
characteristics and demographic characteristics of the housing wealth effect.
ln𝐶𝑖 = 𝛼0 + 𝛼1 ln 𝐻𝑖 + 𝛼2 ln 𝑌𝑖 + ∑ 𝛼𝑘,𝑖𝑋𝑘,𝑖𝑚𝑘=1 + ∑ 𝛼𝑗,𝑖𝑋𝑗,𝑖 ln 𝐻𝑖
𝑙𝑗=1 + 𝜀𝑖 (16)
where j ≠ k.
4. Macro evidence of housing wealth effect
4.1 Sample data
All data of macro part is from the National Bureau of Statistics of China
<http://data.stats.gov.cn/> on a quarterly frequency, from 2006q1 to 2018q2, 50 periods in
total. The consumption and income of residents were measured by the per capita consumption
expenditure of urban residents and the per capita disposable income of urban residents. The
price level of commercial housing is defined as the result of aggregate quarterly sales of
commercial housing divided by the quarterly sales area. The consumption expenditure per
capita is the independent variable; the disposable income per capita and the price level of
commercial housing are used as dependent variables. Logarithmic processing is used to the
19
original data to avoid the heteroscedasticity of the time series. Furthermore, the X12 seasonal
adjustment method in Eviews is used to get rid of the instability caused by the seasonal
variation of the time series.
Table 4.1.1: Descriptive statistics of variables in log-differences (2006Q1 to 2018Q2)
DLNCON1 DLNCON2 DLNINC1 DLNINC2 DLNHP
Mean 0.0232 0.0239 0.0253 0.0264 0.0199
Median 0.0206 0.0265 0.0223 0.0242 0.0144
Maximum 0.0971 0.0900 0.0749 0.2353 0.0970
Minimum -0.0350 -0.2249 -0.0166 -0.1151 -0.0645
Std. Dev. 0.0190 0.0455 0.0146 0.0471 0.0346
Observations 49 49 49 49 49
20
Figure 4.1.1: Line chart of variables in log-differences (2006Q1 to 2018Q2)
-.3
-.2
-.1
.0
.1
.2
.3
06 07 08 09 10 11 12 13 14 15 16 17 18
DLNCON1 DLNCON2 DLNINC1DLNINC2 DLNHP
Year
Va
lue
4.2 General tests and estimation
In this section, the housing wealth effect will be tested with macro evidence of China, from
2006Q1 to 2018Q2. In the following models, consumption is set as the dependent variable,
income and house price are set as independent variables. The purpose of this section is to find
how the change of house price affects consumption in China.
4.2.1 Unit root tests
If the time series has a unit root, it will cause a problem of spurious regression for variables.
Through cointegration analysis, we can investigate the long-term equilibrium relationship
between economic variables, and one of the requirements of cointegration analysis is that the
21
sequence is in the same order process, so it is needed to test the stationary of the original data
and the first-order difference sequence data to verify the applicability of the cointegration
analysis. Only when the sequence is in the same order process cointegration can be used.
Table 4.2.1: Unit root tests (ADF test) (2006Q1 to 2018Q2)
Original
1%
level
5%
level
10%
level
ADF test
statistic Prob.
Consumption of Urban
Residents (lncon1) -3.5713 -2.9224 -2.5992 -1.7889 0.3816
Income of Urban Residents
(lninc1) -3.5713 -2.9224 -2.5992 -2.9657 0.0543
Consumption of Rural
Residents (lncon2) -3.5713 -2.9224 -2.5992 -1.0506 0.7277
Income of Rural Residents
(lninc2) -3.5713 -2.9224 -2.5992 -1.4877 0.5314
House Price (lnhp) -3.5713 -2.9224 -2.5992 -0.9864 0.7511
After first order difference
1%
level
5%
level
10%
level
ADF test
statistic Prob.
Consumption of Urban
Residents (lncon1) -3.5744 -2.9238 -2.5999 -11.8534 0.0000
Income of Urban Residents
(lninc1) -3.5744 -2.9238 -2.5999 -7.6141 0.0000
Consumption of Rural
Residents (lncon2) -3.5744 -2.9238 -2.5999 -6.5585 0.0000
Income of Rural Residents
(lninc2) -3.5744 -2.9238 -2.5999 -9.4513 0.0000
House Price (lnhp) -3.5744 -2.9238 -2.5999 -8.3494 0.0000
The lag selection of the ADF test is based on Schwartz Information Criterion. The ADF tests
are performed on the basis of 5 percent significance level with the null hypothesis that is of no
stationarity. As can be seen in the tables above, all variables are not stationary at conventional
levels at 5% significance level while the null hypothesis is rejected at first differences. So, it
22
is concluded that results indicate that all variables are integrated of order one i.e. I(1), which
allows us to proceed with the cointegration test. Therefore, we can use the method of
Johansen cointegration test to analyze the housing wealth effect of urban and rural residents.
4.2.2 Cointegration tests
Table 4.2.2: Unrestricted Cointegration Rank Test (Trace) of Model(1)
(2006Q1 to 2018Q2)
Hypothesized
No. of CE(s)
Eigenvalue Trace Statistic 0.05 Critical
Value
Prob.
None * 0.4091 47.9971 29.7971 0.0002
At most 1 * 0.2374 23.2677 15.4947 0.0028
At most 2 * 0.2007 10.5302 3.8415 0.0012
Trace test indicates 3 cointegrating equations at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
Table 4.2.3: Unrestricted Cointegration Rank Test (Trace) of Model (2)
(2006Q1 to 2018Q2)
Hypothesized
No. of CE(s)
Eigenvalue Trace Statistic 0.05 Critical
Value
Prob.
None * 0.4091 47.9971 29.7971 0.0002
At most 1 * 0.2374 23.2677 15.4947 0.0028
At most 2 * 0.2007 10.5302 3.8415 0.0012
Trace test indicates 3 cointegrating equations at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
The test results are listed in table 4.2.2 and table 4.2.3, noted as model (1), which represents
the urban model, and model (2), which is the rural model. Trace test indicates that there are
more than 2 cointegration equations at the 0.01 level of significance for model (1) and model
(2).
4.2.3 Collinearity autocorrelation and heteroskedasticity
We can see that the correlation coefficients among variables are low, as shown in table 4.2.4
and table 4.2.5, so we can exclude multicollinearity among the explanatory variables.
23
Through OLS we can get DW statistics of each equation. The results of DW test are all
d𝑢 < DW < 4 − d𝑢, so there is no autocorrelation. However, White test results (shown in
table 4.2.6) show that there is heteroskedasticity in OLS models, so Weighted Least Squares
(WLS) are used instead of OLS in order to reduce the effects of heteroscedasticity. Minimize
∑ 𝑤𝑖(𝑦𝑖 − 𝛼0 − 𝛼1𝑥𝑖1 − ⋯ − 𝛼𝑝𝑥𝑖𝑝)2𝑛𝑖=1 , where 𝑤𝑖 =
1
𝜎𝑖2 , then get the estimation of 𝛼. The
estimation results are listed in table 4.2.8.
To get the estimates of cointegrating equations, Fully Modified Least Squares (FMOLS)
regression is introduced here and results are listed in table 4.2.7. FMOLS was originally
designed in work by Phillips and Hansen (1990) to provide optimal estimates of cointegrating
regressions. This method modifies OLS to account for serial correlation effects and for the
endogeneity in the regressors that result from the existence of a cointegrating relationship.
Plus, compared with Maximum Likelihood Estimate, FMOLS is better in dealing with small
sample cointegrating regressions.
Table 4.2.4: The correlation among variables of model (1) (2006Q1 to 2018Q2)
LNCON1 LNINC1 LNHP
LNCON1 1.0000 0.4393 0.1956
LNINC1 0.4393 1.0000 -0.1999
LNHP 0.1956 -0.1999 1.0000
DLNCON1 DLNINC1 DLNHP
DLNCON1 1.0000 0.6020 0.3063
DLNINC1 0.6020 1.0000 -0.0951
DLNHP 0.3063 -0.0951 1.0000
24
Table 4.2.5: The correlation among variables of model (2) (2006Q1 to 2018Q2)
LNCON2 LNINC2 LNHP
LNCON2 1.0000 0.2598 -0.0570
LNINC2 0.2598 1.0000 -0.1284
LNHP -0.0570 -0.1284 1.0000
DLNCON2 DLNINC2 DLNHP
DLNCON2 1.0000 0.2846 -0.0870
DLNINC2 0.2846 1.0000 -0.1900
DLNHP -0.0870 -0.1900 1.0000
Table 4.2.6: Heteroskedasticity Test White (2006Q1 to 2018Q2)
White Test of model (1)
F-statistic 7.6502 Prob. F(5,43) 0
Obs*R-squared 23.0681 Prob. Chi-Square(5) 0.0003
Scaled explained SS 36.4494 Prob. Chi-Square(5) 0
White Test of model (2)
F-statistic 13.3299 Prob. F(5,43) 0
Obs*R-squared 29.7842 Prob. Chi-Square(5) 0
Scaled explained SS 149.6246 Prob. Chi-Square(5) 0
4.2.4 Estimation results and analysis
Table 4.2.7: Estimates of cointegration equations (2006Q1 to 2018Q2)
(long run housing wealth effect)
Model(1) Model(2)
Constant 0.1295 0.9672*
(1.2266) (1.8691)
lnhp 0.0633 -0.0271
(1.4481) (-0.1742)
lninc1 0.8803***
(26.8344)
lninc2 0.8850***
(7.8191)
𝐑𝟐 0.9986 0.9733
Method: Fully Modified Least Squares (FMOLS)
Note: *,**,*** denote significance at the 10,5,1 percent level respectively. T-statistics is in parentheses.
25
Long run cointegration equations:
Model (1):
ln 𝐶𝑡 = 0.1295 + 0.0633 ln 𝐻𝑃𝑡 + 0.8803 ln 𝑌𝑡
t: 1.2266 1.4481 26.8344
p-value: 0.2261 0.1542 0.0000
Model (2):
ln 𝐶𝑡 = 0.9672 − 0.0271 ln 𝐻𝑃𝑡 + 0.8850 ln 𝑌𝑡
t: 1.8691 -0.1742 7.8191
p-value: 0.0678 0.8625 0.0000
Table 4.2.8: Estimates of short run VECM equations (2006Q1 to 2018Q2)
(short run housing wealth effect)
Model(1) Model(2)
Constant -0.0088*** 0.0276***
(-2.9851) (5.4414)
dlnhp 0.1467*** 0.1005
(3.6257) (0.8973)
dlninc1 1.1690***
(12.2193)
dlninc2 0.2681*
(1.9327)
𝐑𝟐 0.3006 0.0162
𝐑𝒘𝟐 0.8084 0.0318
𝐑𝒏𝟐 -statistic 155.4348 1.1205
Method: WLS
Note: *,**,*** denote significance at the 10,5,1 percent level respectively. Z-statistics is in parentheses.
Short run VECM equations:
Model (1):
26
dln 𝐶𝑡 = −0.0088 + 0.1467 dln 𝐻𝑃𝑡 + 1.1690 dln 𝑌𝑡
z: -2.9851 3.6257 12.2193
p-value: 0.0028 0.0003 0.0000
Model (2):
dln 𝐶𝑡 = 0.0276 + 0.1005 dln 𝐻𝑃𝑡 + 0.2681 dln 𝑌𝑡
z: 5.4414 0.8973 1.9327
p-value: 0.0000 0.3695 0.0594
The results of equations above show that, overall, there is a positive relationship between
house price and consumption in the urban area in China while that relationship in the rural
area is not significant. In detail, in the model (1) which represents urban area, for every 1%
increase in the housing price, consumption will increase by 0.0633% (though t-statistics
shows it is not very significant with a p-value of 0.1542) in the long run and consumption will
increase by 0.2012% (which is a very significant number with a t-statistics of 3.4888) in the
short run; while for model (2), the relationship between house price and consumption is
slightly negative and not significant in both long run and short run, which means that the
change of house price has little impact on rural households’ consumption. This finding is
consistent with direct housing wealth effect and also shows a significant difference of urban
area and rural area.
In addition, the goodness of fit (R2) of the two cointegration equations are over 0.95 and the
27
goodness of fit of the short run VECM equations are less than 0.5, especially for the model (2)
with a number of 0.0821. This is to say, in the long run, income and house price could explain
over 95% changes of consumption while in the short run, this number declines to 49.58% for
urban area and 8.21% for rural area respectively. Meanwhile, the F-statistics of urban area are
larger than that of rural area in both long run and short run.
The data and figures of impulse responses and variance decomposition between consumption
and house price are listed in the appendix. For model (1), an unexpected positive shock to
house price will lead to continuous growth of consumption in 10 quarters’ period, and the
accumulated change is about 0.05% for every 1% shock to house price. In contrast, the
impulse response of consumption to house price in model (2) is weaker, where the 10 quarters’
accumulated response is 0.035%. The difference of variance decomposition results is much
bigger where in mode (1) house price could explain about 40% variance of consumption but
this number of model (2) is only 0.5%. Therefore, the results of impulse responses and
variance decomposition suggest that the change of house price will make more influence on
consumption in urban area than that in rural area, which is also in accordance with the result
of former equation estimation.
Under the hypothesis that the annual housing area increase is small compared with the stock
area number, the relationship between house price and consumption could be used to imply
the relationship between housing wealth and consumption. So it could be concluded as there
28
is a positive housing wealth effect in urban areas of China and the housing wealth effect in
rural area is not significant.
Table 4.2.9: Quandt-Andrews unknown breakpoint test (2006Q1 to 2018Q2)
Statistic Value Prob.
Maximum LR F-statistic (2013Q1) 13.52683 0
Maximum Wald F-statistic (2013Q1) 40.58049 0
Null Hypothesis: No breakpoints within 15% trimmed data
In order to further analyze the housing wealth effect in urban areas, I observed that the growth
rate of house prices had declined after 2012, while the income and consumption data of urban
residents maintained relatively stable growth. The result of Quandt-Andrews unknown
breakpoint test, as shown in table 4.2.9, shows that 2013q1 is a breakpoint of series .So
consider taking 2013q1 as the breakpoint and dividing the total sample into two sections for
analysis: from the first quarter of 2006 to the fourth quarter of 2012 as section 1 and from the
first quarter of 2013 to the second quarter as section 2, 28 quarters and 22 quarters
respectively.
4.3 Subsample analysis (2006Q1–2012Q4)
Take ADF test to each variable in the first section under the null hypothesis that the time
series has a unit root and is non-stationary while under the alternative hypothesis this is not
the case, the test results (Table 4.3.1) show that the original sequences of consumption,
income and housing price level have unit roots, while the first order difference sequences
remain stationary under the significance level of 1%. So we can get the conclusion that the
logarithmic sequences of per capita consumption, income and housing price level after
29
seasonal adjustment are first-order differential stationary .
Table 4.3.1: Unit root test (ADF test) (2006Q1 to 2012Q4)
Original
1% level 5% level 10% level ADF test statistic Prob.
Consumption -3.6999 -2.9763 -2.6274 -0.7642 0.8132
Income -3.6999 -2.9763 -2.6274 -0.7821 0.8082
House Price -3.6999 -2.9763 -2.6274 -0.9406 0.7592
After first order difference
1% level 5% level 10% level ADF test statistic Prob.
Consumption -3.7115 -2.9810 -2.6299 -12.5121 0.0000
Income -3.7115 -2.9810 -2.6299 -6.5597 0.0000
House Price -3.7115 -2.9810 -2.6299 -5.2544 0.0002
In order to obtain the optimal estimation model, I made some adjustments to the original
model, and got four models as shown in table 4.3.2. The original estimation model 𝑐𝑡 =
𝛼0ℎ𝑝𝑡 + 𝛼1𝑦𝑡 + 𝜀𝑡 is model (1), and the Model (2) is obtained after adding a constant term.
The results show that although the constant term is only significant at 10% level, the goodness
of fit (R2) improves from 0.61 to 0.67. The significance of elasticity of consumption spending
to changes in house price improves from non-significant to significant at level of 5%, while
income remains very significant. Model (2) shows that for every 1% increase in income, the
consumption will increase by 1.16%, that is, MPC is 1.16, larger than 1; for every 1%
increase in house price, the consumption will increase by 0.15%, that is, the real estate wealth
effect is estimated to be 0.15. Compared with model (1), after adding a constant term, R2 and
the significance of dlnhp of model (2) improves. But with the result that MPC>1, it means
30
that the marginal saving propensity (MPS) is negative. The permanent income hypothesis
holds that consumption is smooth and that the impact of temporary income changes on
consumption will be distributed to the entire life cycle, which means that it has little influence
on the current period. So, considering the result that MPC is 1.16 of model (2), model (1) is
more compatible with the traditional economics theory.
In order to improve the overall statistical characteristics of the equation, I try to add long-term
equilibrium factors and estimate the long run housing wealth effect in the meantime. Since
that the consumption, housing price level and income of urban residents are all first-order
differential stationary sequences, I try to establish a long-term equilibrium relationship
equation:
W = lnC𝑡 − c − β0 lnHP𝑡 − β1 lnY𝑡 (14)
the result of which is shown in table 4.3.3. As shown in table 4.3.4, W is stationary, which
means that there is cointegration relationship among income, consumption and housing price.
Since the long-term equilibrium relationship established is established, error correction model
is applicable as the estimation model.
After adding a long-term equilibrium relationship W, the model (3) and model (4) are
obtained. The results show that there is a significant improvement in R2 and DW. After
adding W, the model (1) changes to the model (3), with the result that R2 improve from 0.61
to 0.85; the model (2) changes to the model (4), with the result that R2 improve from 0.67 to
31
0.87. In terms of statistical characteristics and economic meanings, the long-term equilibrium
relation equation model (3) with no constant is better. According to estimation result of W, the
consumption spending of residents mainly depends on their income level. In the long run,
residents spend 0.81% more on consumption out of every 1% of increase in income, while
that of long-term housing price level increase is smaller with a number of only 0.11%.
According to the results of model (3), for every 1% increase in housing price, consumption
will increase by 0.14%, and for every 1% increase in income, consumption will increase by
0.83%. Combined with long-term factors and short-term factors, the change of housing prices
will cause the change of residents’ consumption, and its short-term effect is slightly larger
than the long-term effect.
Table 4.3.2: Estimates of short run coefficients (2006Q1 to 2012Q4)
Model(1) Model(2) Model(3) Model(4)
Constant -0.0127* -0.0091**
(-2.0348) (-2.2886)
dlnhp 0.1120 0.1523** 0.1357*** 0.1632***
(-1.5139) (-2.1007) (2.8476) (3.5833)
dlny 0.8563*** 1.1626*** 0.8282*** 1.0484***
(-9.0854) (-6.6511) (13.6508) (9.4201)
W 1.1692*** 1.1079***
(6.0564) (6.1555)
𝐑𝟐 0.6103 0.6677 0.8459 0.8745
DW 3.2933 3.0678 1.8690 1.8822
Note: *,**,*** denote significance at the 10,5,1 percent level respectively.
32
Table 4.3.3: Estimates of long run coefficients (2006Q1 to 2012Q4)
W
Constant 0.3327***
(2.9880)
lnHP 0.1105**
(2.6490)
lnY 0.8080***
(24.9696)
𝐑𝟐 0.9980
Note: *,**,*** denote significance at the 10,5,1 percent level respectively.
Table 4.3.4: Unit root test of W (2006Q1 to 2012Q4)
1% level 5% level 10% level ADF test statistic Prob.
W -3.6999 -2.9763 -2.6274 -5.7736 0.0001
Results analysis
1. Short-term housing wealth effect is consistent with the findings. The empirical test results
show that in the short term, for every 1% increase in housing price, consumption will increase
by 0.14%, which indicates that there is a positive short-term housing wealth effect of 0.14,
significant at 1% level.
2. Long-term housing wealth effect is also found to be consistent with the evidence. There is a
cointegration relationship among income, consumption and housing price. In the long term,
for every 1% increase in housing price, consumption will increase by 0.11%, which indicates
that there is a positive short-term housing wealth effect of 0.11, significant at 5% level.
3. Short-term housing wealth effect is larger than long-term housing wealth effect with a gap
33
of 0.03.
4. The impact of household income on consumption is greater, both in the long and short term.
The impact of house price changes on consumption is less than 0.2, much smaller than that of
income changes on consumption with a number of more than 0.8. This shows that income is
the determining factor of consumption level, which is consistent with the view of traditional
consumption theory.
4.4 Subsample analysis (2013Q1–2018Q2)
The ADF test results (Table 4.4.1) are similar to the data of the first period. The logarithm of
variables are first-order differential stationary after seasonal adjustment.
Table 4.4.1: Unit root test (ADF test) (2013Q1 to 2018Q2)
Level
1% level 5% level 10% level ADF test statistic Prob.
Consumption -3.7880 -3.0124 -2.6461 -1.0931 0.6985
Income -3.8085 -3.0207 -2.6504 1.2890 0.9975
House Price -3.8085 -3.0207 -2.6504 -1.9305 0.3126
First difference
1% level 5% level 10% level ADF test statistic Prob.
Consumption -3.8085 -3.0207 -2.6504 -4.2644 0.0038
Income -3.8085 -3.0207 -2.6504 -6.4752 0.0000
House Price -3.8085 -3.0207 -2.6504 -6.8265 0.0000
In contrast to the previous results, no matter what adjustment done to the model form, the test
results (Table 4.4.2) show that the coefficient of housing price changes is very small and not
34
significant. Meanwhile, the overall goodness of fit is also very poor. In the long-term
equilibrium relationship (Table 4.4.3), the long-term housing wealth effect is also not
significant, while the impact of income on consumption is very significant with a coefficient
over 0.9. The results show that during this period, whether in the long or short term, the
change of income is the main factor to explain consumption change, while the impact of
house price changes on consumption is minimal.
Table 4.4.2: Estimates of short run coefficients (2013Q1 to 2018Q2)
Model(1) Model(2) Model(3)
Constant 0.0193
1.8936
dlnhp 0.0587 -0.0054
0.7378 -0.0654
dlny 0.9166*** 0.1306 0.9474***
8.4806 0.3057 9.6125
W
R2 -0.1911 0.0068 -0.2252
DW 2.3893 1.8002 2.4633
Note: *,**,*** denote significance at the 10,5,1 percent level respectively.
35
Table 4.4.3: Estimates of long run coefficients (2013Q1 to 2018Q2)
W1 W2 W3
Constant 0.0476
0.2698
lnHP 0.0118 0.0232
0.2182 0.6971
lnY 0.9434*** 0.9372*** 0.9610***
22.6197 27.4655 4882.2680
𝐑𝟐 0.9965 0.9965 0.9964
DW 1.6308 1.6122 1.6037
Note: *,**,*** denote significance at the 10,5,1 percent level respectively.
During this period, the change of income is the main factor to explain consumption change,
while the impact of house price changes on consumption is minimal, indicating that the
housing wealth effect is not significant. There are four possible reasons. First, because of the
small amount of sample data, the results are less credible. Second, because of the impact of
the price-control policy, the nominal housing price deviated from the real value, and
consumers expect this policy will not last long, so they believe there is little impact on
household wealth, which is also consistent with life cycle hypothesis. Third, the real estate
wealth effect has a directional characteristic, that is, significant with positive changes and not
significant with negative changes, indicating that consumers will increase consumption
spending when house prices rise, but will not react when house prices decline. Finally, it is
possible that the housing wealth effect in China has indeed decreasing in recent years.
36
5. Micro evidence of housing wealth effect
5.1 Sample data and tests
In order to further investigate the reasons for the weakening of the housing wealth effect in
recent years and the transmission mechanism, I use latest survey data of CHNS(CHNS2009,
CHNS2011 and CHNS2015) <https://www.cpc.unc.edu/projects/china/data/datasets> to
examine the impact of house value on household consumption. Data descriptive statistics is
listed in table 5.1.1 (see in appendix). The China Health and Nutrition Survey (CHNS) is an
international collaborative project between the Carolina Population Center at the University
of North Carolina at Chapel Hill and the National Institute for Nutrition and Health (NINH,
former National Institute of Nutrition and Food Safety) at the Chinese Center for Disease
Control and Prevention (CCDC). This survey was designed to collect the data of the health,
nutrition and economic status of Chinese family to see how the social and economic
transformation of Chinese society is affecting the health and nutritional status of its
population.
The use of microscopic household survey data can not only avoid the shortcomings of using
macro data, which is that the sample size of time series data of China is still small, reducing
the credibility of the estimation results. The household survey data with large sample size and
detailed index category, has obvious advantages compared to macro data, especially in
revealing the mechanism of housing wealth effect. We can analyze the differences of different
37
regions, different income layers and different age groups in detail, so as to greatly improve the
accuracy of the analysis and obtain more reliable, more detailed and more persuasive
empirical results.
The correlation coefficients (see in table 5.1.2 in appendix) among the most explanatory
variables are low. Therefore, it is verified that there is no collinearity among the explanatory
variables. Using the White test in Stata, there is heteroskedasticity among errors at 5% level of
significance. Then the robust least square regression is used.
Table 5.1.3: White's test
White's test for Ho: homoskedasticity
chi2 p-value
Heteroskedasticity 935.66 0
5.2 Estimation of equations and analysis
The empirical model is equation (17). ln𝐶𝑖 , ln 𝐻𝑖 , ln 𝑌𝑖 represents the logarithm of the
consumption, house value and household income of family i, respectively. X represents
control variables, including the age, gender, type of work, level of education, family size,
survey time, region, etc. Among them, “hhsize” is the number of family members; “urban”
represents urban-rural difference, where 1 denotes urban area and 0 denotes rural area;
“gender” represents gender, where 0 denotes male and 1 denotes female; “educ” represents
the highest level of education attained, where 0 denotes none,1 denotes primary school, 2
denotes lower middle school, 3 denotes upper middle school,4 denotes technical or vocational
degree, 5 denotes the university or college degree, 6 denotes the master's degree or higher;
38
“mortgage” represents household mortgage condition, where 0 denotes no mortgage and 1
denotes that the family have at least one mortgage; the years of the survey are 2009, 2011 and
2015, where 2009 years as the base period, “t2011” equals 1 when the survey year is 2011 and
“t2015” equals 1 when the survey year is 2015; the survey area is divided into 4 parts,
northeastern, eastern, central and western China respectively, with northeastern China as the
base. 𝑋𝑗,𝑖 ln 𝐻𝑖 denotes the combined term of virtual variables and the log of house value.
ln𝐶𝑖 = 𝛼0 + 𝛼1 ln 𝐻𝑖 + 𝛼2 ln 𝑌𝑖 + ∑ 𝛼𝑘,𝑖𝑋𝑘,𝑖𝑚𝑘=1 + ∑ 𝛼𝑗,𝑖𝑋𝑗,𝑖 ln 𝐻𝑖
𝑙𝑗=1 + 𝜀𝑖 (17)
where j≠k.
By estimating the coefficient of the combined term, we can further test the housing wealth
effect in different genders, different ages, different years and different regions. By estimating
𝛼1, we can get the elasticity of household consumption expenditure to changes in housing
wealth, and examine housing wealth effect on the microscopic. The test results are listed in
table 5.2 (see in appendix).
Model (1) shows that household income and house value have a significant impact on
household consumption. The elasticity of consumption to changes in income is 0.63, which
means that for every 1% increase in household income, the consumption will increase by
0.63%; the elasticity of consumption to changes in house value is 0.07, which means that for
every 1% increase in home value, the consumption will increase by 0.07%. This result is in
accordance with the direct wealth effect that residents will increase consumption when their
houses are worth more money.
39
Model (2) shows that among the control variables, the coefficients of “urban”, “age” and
“educ” are very significant (1% level); the coefficient of “mortgage” is significant at 5% level;
the coefficient of “gender” is significant at the level of 10%; “hhsize” is not significant.
Compared with rural households, urban household consumption expenditure is about 50%
larger. The consumption expenditure of men is about 6% higher than that of women. With the
increase of age and education, the people’s saving rate is increasing and the consumption rate
decreased slightly. The consumption expenditure of households with mortgage is about 22%
lower than households with no mortgage on average.
Model (3) ~ (6) examine the change of housing wealth effect under different conditions. For
urban households, the elasticity of consumption spending to changes in house wealth is about
0.10, which is obviously higher than 0.05 of rural households; the household mortgage has no
significant impact on the housing wealth effect. For men, the elasticity of consumption to
changes in house value is slightly higher than that of women. Of the regional factors, the
housing wealth effect is the most significant in northeast China, followed by the western,
central and eastern China in order. In terms of time factors, the elasticity of consumption
spending to changes in housing wealth shows a gradual decline. Compared with 2009, the
number of housing wealth effect in 2011 decreases by 0.02, and it decreased by 0.09 until
2015, which is to say the housing wealth effect is on a downward trend in recent years and it
even became a negative number in 2015.
40
5.3 Robustness test: house owners and the houseless
There are 210 pieces of data provided by houseless people in the total sample with a sample
size of 9231. So if we take a contrast of house owners and the houseless, we can find that
there is a totally different way on how housing wealth effect functions.
Table 5.3: Estimates of house owners and the houseless
total sample house owners the houseless
R2 0.1808 0.1815 0.1418
constant 1.1061*** 1.0289*** 4.8804**
5.9 4.2 3.98
lnY 0.6310*** 0.6365*** 0.3402***
38.74 28.61 3.41
lnH 0.0589*** 0.0603*** -0.2634
15.5 14.38 -0.95
hhsize -0.0023 -0.0028 0.0479
-0.23 -0.28 0.65
urban 0.5037*** 0.5097*** -1.5533***
9.43 8.05 -5.12
gender -0.0647* -0.0628** 0.1235
-1.95 -2.04 0.61
age -0.0063*** -0.0062*** -0.0154**
-6.31 -6.1 -2.19
educ -0.1023*** -0.1015*** -0.1908*
-6.74 -6.07 -1.94
mortgage -0.2184** -0.2186*
-2.18 -1.83
Method: robust least squares
Note: *,**,*** denote significance at the 10,5,1 percent level respectively. T statistics are below the
coefficients.
There are two major differences between house owners and the houseless. First, for every 1%
positive shocks to house value, the house owners will increase around 0.06% consumption
while the houseless tend to cut spending on consumption since they do not own houses and
when the house appreciates, they will have to pay more rents. Another difference is that
41
people who live in cities without houses spend much less on consumption compared with the
houseless living in the countryside, which is on the opposite with house owners. A possible
reason may be that rents in cities are much higher than that in countryside, crowding out quite
an amount of consumption, which is consistent with the budget constraint effect as mentioned
before. When 210 pieces of houseless data are removed from the data sample, the estimation
results change little compared with the original sample estimates, so we can say the former
results are robust.
6. Conclusion
To some extent, the results of micro models in this paper are consistent with that of macro
models. The overall housing wealth effect in China is around 0.06 and there is a difference as
large as 0.05 between town and countryside. The micro evidence shows that the housing
wealth effect on household consumption is generally positive, differing from regions, years,
gender, education and other factors. The micro evidence also finds that for people who own
houses and people who are houseless, the impact of a shock to the housing assets value on
consumption could be completely opposite.
After 2013, the housing wealth effect shows a trend of weakening in recent years, especially
in 2015, when the elasticity of consumption to changes in housing wealth is negative. Possible
reasons of the decreasing housing wealth effect may include: first, the sample amount after
2013 is quite small, which could make the results are less credible; second, for the sake of
government intervention, the housing price is deviating from the market price, and consumers
42
expect this policy will not have a long-term impact on their wealth; third, the housing wealth
effect is limited compared with income effect, so when people’s income is growing at a high
speed, consumption will keep growing even their assets do not appreciate; finally, it is
possible that the consumption habit is changing in recent years because the Internet finance is
popular in China and more people begin to accept credit consumption when can get easy
credit at online shopping.
43
Reference
Alexander D. UK Government: Alexander challenges business–“Social responsibility must
not be just skin deep”[J]. Coventry: M2 Presswire, 2002.
Ando A, Modigliani F. The" life cycle" hypothesis of saving: Aggregate implications and
tests[J]. The American economic review, 1963, 53(1): 55-84.
Bayoumi T, Edison H. Is wealth increasingly driving consumption?[M]. De Nederlandsche
Bank, 2003.
Carroll C D, Otsuka M, Slacalek J. How large are housing and financial wealth effects? A new
approach[J]. Journal of Money, Credit and Banking, 2011, 43(1): 55-79.
Case K E, Quigley J M, Shiller R J. Comparing wealth effects: the stock market versus the
housing market[J]. Advances in macroeconomics, 2005, 5(1).
Catte P, Girouard N, Price R W R, et al. Housing markets, wealth and the business cycle[J].
2004.
Cheney J S. Supply-and demand-side developments influencing growth in the debit market[R].
Federal Reserve Bank of Philadelphia, 2006.
Cheng A C S, Fung M K. Financial market and housing wealth effects on consumption: a
permanent income approach[J]. Applied Economics, 2008, 40(23): 3029-3038.
Carroll C D, Crawley E, Slacalek J, et al. Sticky expectations and consumption dynamics[R].
National Bureau of Economic Research, 2018.
Díaz A, Luengo‐Prado M J. The wealth distribution with durable goods[J]. International
Economic Review, 2010, 51(1): 143-170.
Ding Pan, Hu Zongyi. Dynamic study on the influence of stock price and house price
44
fluctuation on residents' consumption [J]. Statistics and Decision-making, 2008 (15): 106-108.
Dvornak N, Kohler M. Housing wealth, stock market wealth and consumption: a panel
analysis for Australia[J]. Economic Record, 2007, 83(261): 117-130.
Elliott J W. Wealth and wealth proxies in a permanent income model[J]. The Quarterly
Journal of Economics, 1980, 95(3): 509-535.
Jin Tao, Jiang Kai. A Study on the relationship between Chinese residents ' consumption and
house price from the perspective of generalized virtual economy {J}. Generalized Virtual
Economy Study, 2012 (03): 35-41.
Kiyotaki N, Michaelides A, Nikolov K. Winners and losers in housing markets[J]. Journal of
Money, Credit and Banking, 2011, 43(2‐3): 255-296.
Laurence L. Are assets fungible? testing the behavioral theory of Life-Cycle savings[J].
Journal of Economic Organization and Behavior, 1998, 36: 59-83.
Li Yaming, Tong Rencheng. Cointegration analysis and error correction model of housing
wealth effect in China [J]. Theory and practice of System Engineering, 2007 (11): 1-7.
Liu Dan. Analysis of wealth effect of urban residential market in China--macroscopic
consumption function based on life cycle hypothesis [J]. Journal of the Capital Economic and
Trade University, 2007 (04): 108-112.
Liu Jianjiang, Yang Yujuan, Yuan Dongmei. The function mechanism of housing wealth effect
from the perspective of consumption function theory [J]. Consumer Economy，2005 (02)：
93-96
Liu Yufei, Zhou Yingjie, Yang Zhengyu. Housing wealth value and upgrading of household
consumption structure - evidence from CFPS (2014) data [J]. Journal of Hebei University of
Economics and Trade, 2018, 39 (4):22-34.
Ludwig A, Sløk T. The impact of changes in stock prices and house prices on consumption in
45
OECD countries[M]. International Monetary Fund, 2002.
Luo Zuoyan. An analysis of the wealth effect of urban residents’ housing assets based on
liquidity [J]. Contemporary Economic Science, 2007, 29 (4): 51-56.
Modigliani F, Brumberg R. Utility analysis and the consumption function: An interpretation of
cross-section data[J]. Franco Modigliani, 1954, 1: 388-436.
Muellbauer J. Housing and personal wealth in a global context[M]. Research Paper,
UNU-WIDER, United Nations University (UNU), 2007.
Peek J. Capital gains and personal saving behavior[J]. Journal of Money, Credit and Banking,
1983, 15(1): 1-23.
Phillips P C B, Hansen B E. Statistical inference in instrumental variables regression with I (1)
processes[J]. The Review of Economic Studies, 1990, 57(1): 99-125.
Sierminska E, Takhtamanova Y. Disentangling the wealth effect: some international
evidence[J]. FRBSF Economic Letter, 2007.
Tang Jianwei. The impact of changes in stock and real estate prices on consumption [J].
Journal of Shanghai Institute of Finance, 2004, 1:48-51.
Wei Feng. Wealth effects in China's stock and housing markets [J]. Journal of Chongqing
University (Natural Science Edition), 2007 (02): 153-157.
Zhaoxiaoli, Ma Hui, Chen Shoudong. The influence of stock price and house price on
consumption behavior [J]. Journal of Guangzhou University (Social Science Edition), 2007
(3): 48-53
46
Appendix
Table：Impulse response and variance decomposition
lncon1 lncon2
Periods Impulse
response
Accumulated
response
Variance
decomposition (%)
Impulse
response
Accumulated
response
Variance
decomposition (%)
1 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
2 0.0010 0.0010 0.4306 -0.0003 -0.0003 0.0024
3 0.0009 0.0020 0.5586 0.0029 0.0025 0.0993
4 0.0033 0.0052 3.0307 0.0036 0.0062 0.1811
5 0.0047 0.0099 7.1191 0.0049 0.0111 0.3001
6 0.0058 0.0157 12.0287 0.0047 0.0158 0.3622
7 0.0072 0.0229 18.0651 0.0050 0.0208 0.4159
8 0.0082 0.0310 24.1700 0.0050 0.0257 0.4552
9 0.0091 0.0402 30.1869 0.0049 0.0306 0.4822
10 0.0100 0.0502 35.8860 0.0049 0.0355 0.5044
47
Figure 4.2.1 to 4.2.6
-.005
.000
.005
.010
.015
.020
1 2 3 4 5 6 7 8 9 10
Response of LNCON1 to LNHPChange of lncon1
Periods
-.04
-.02
.00
.02
.04
.06
.08
1 2 3 4 5 6 7 8 9 10
Accumulated Response of LNCON1 to LNHPChange of lncon1
Periods
48
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10
Percent LNCON1 variance due to LNHPPercent of lncon1 variance
Periods
-.02
-.01
.00
.01
.02
.03
.04
.05
.06
1 2 3 4 5 6 7 8 9 10
Response of LNCON2 to LNHPChange of lncon2
Periods
49
-.2
-.1
.0
.1
.2
.3
.4
.5
.6
1 2 3 4 5 6 7 8 9 10
Accumulated Response of LNCON2 to LNHPChange of lncon2
Periods
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10
Percent LNCON2 variance due to LNHPPercent change of lncon2 variance
Periods
50
Table 5.1.1: Descriptive statistics of transformed CHNS data
Variable Dscription n Mean Std.Dev. Min Max
lnC log of yearly
consumption
9231 7.89 1.74 1.90 14.11
lnY log of yearly income 9231 10.69 1.06 0 14.24
lnH log of housing assets
value
9231 8.10 4.51 0 16.20
hhsize num. of family
members
9231 4.04 1.68 1 15
urban living in the urban
area
9231 0.11 0.32 0 1
gender 0 for male, 1 for
female
9231 0.49 0.50 0 1
age age 9231 44.21 17.34 4 100
education the highest level of
education attained
9231 1.71 1.17 0 6
mortage 1 for mortage 9231 0.03 0.16 0 1
t2011 1 for year 2011 9231 0.34 0.47 0 1
t2015 1 for year 2015 9231 0.42 0.49 0 1
east 1 for living in the
east
9231 0.15 0.36 0 1
mid 1 for living in the
mid
9231 0.32 0.47 0 1
west 1 for living in the
west
9231 0.26 0.44 0 1
51
Table 5.1.2: Correlation among variables
lnC lnY lnH hhsize urban gender age education
lnC 1.0000
lnY 0.3678 1.0000
lnH 0.1437 -0.0950 1.0000
hhsize 0.0652 0.1672 -0.0198 1.0000
urban 0.1395 0.0993 0.1277 -0.0497 1.0000
gender -0.0089 -0.0230 0.0529 0.0312 0.0046 1.0000
age -0.0939 -0.1048 -0.0380 -0.2308 -0.0053 -0.0207 1.0000
education 0.0110 0.2309 -0.2178 0.0347 0.1183 -0.1294 -0.1941 1.0000
east -0.0110 0.0991 -0.0204 -0.0399 0.0562 -0.0075 0.0540 0.1006
mid -0.0415 -0.0018 -0.0309 0.1486 -0.0150 -0.0025 -0.0231 0.0414
west -0.0409 -0.0717 0.0225 0.1402 0.1130 -0.0014 -0.0509 -0.0860
t2011 0.0522 -0.0680 0.6264 -0.0539 0.0666 0.0448 0.0162 -0.1749
t2015 -0.1247 0.1606 0.9534 0.0647 -0.0697 -0.0538 0.0118 0.2746
east mid west t2011 t2015
lnC
lnY
lnH
hhsize
urban
gender
age
education
east 1.0000
mid -0.2954 1.0000
west -0.2552 -0.4123 1.0000
t2011 0.0472 -0.0938 0.0033 1.0000
t2015 0.0528 0.0315 0.0075 -0.6061 1.0000
52
Table 5.2: Estimates of households’ consumption (total sample)
Model(1) Model(2) Model(3) Model(4) Model(5) Model(6)
R2 0.1675 0.1808 0.1810 0.1803 0.1807 0.1928
constant 0.5482*** 1.1061*** 1.1331*** 1.1098*** 1.0722*** 1.1211***
3.14 5.9 6.03 5.91 5.75 5.93
lnY 0.6337*** 0.6310*** 0.6318*** 0.6302*** 0.6311*** 0.6417***
40.28 38.74 38.81 38.69 38.75 39.16
lnH 0.0694*** 0.0589*** 0.0544*** 0.0586*** 0.0618*** 0.0713***
18.86 15.5 13.95 15.36 14.77 10.61
hhsize -0.0023 -0.0024 -0.0026 -0.0024 0.0220**
-0.23 -0.24 -0.25 -0.24 2.09
urban 0.5037*** 0.4949*** 0.5031*** 0.6041***
9.43 9.25 9.42 11.12
gender -0.0647* -0.0628* -0.0650** -0.0634*
-1.95 -1.89 -1.96 -1.93
age -0.0063*** -0.0062*** -0.0062*** -0.0062*** -0.0058***
-6.31 -6.25 -6.23 -6.26 -5.82
educ -0.1023*** -0.1013*** -0.1029*** -0.1020*** -0.0894***
-6.74 -6.68 -6.77 -6.72 -5.85
mortgage -0.2184** -0.2354** -0.2188** -0.1919*
-2.18 -2.34 -2.18 -1.91
lnH*urban 0.0490***
9.60
lnH*mortgage 0.0000
0.995
lnH*gender -0.0062*
-1.75
lnH*east -0.0375***
-6.47
53
lnH*mid -0.0314***
-6.60
lnH*west -0.0267***
-5.32
lnH*t2011 -0.0199***
-5.27
lnH*t2015 -0.0946***
-6.83
Method: robust least squares
Note: *,**,*** denote significance at the 10,5,1 percent level respectively. T statistics are below the
coefficients.
